A new application for retail portfolio management has emerged, utilizing personalized, tax-aware reinforcement learning to enhance individual investor capabilities. Unlike existing robo-advisors, which predominantly rely on static, rule-based systems, this application integrates a FastAPI backend with a web dashboard that allows users to voice their investment goals in natural language. For instance, a user could specify a desire for steady growth while needing to liquidate some assets for an upcoming down payment.
The application employs a sophisticated three-phase reinforcement learning architecture. This includes a self-supervised cross-asset encoder, a Mixture-of-Experts (MoE) allocation policy with a learned intent router, and a lightweight Low-Rank Adaptation (LoRA) adapter. These components collaborate to deliver personalized investment recommendations based on each user’s historical brokerage behavior without necessitating retraining of the broader model.
Functionally, the system has been developed, integration-tested, and is currently validated against a live brokerage API, specifically Alpaca, in a paper-trading mode. Key features include multi-user authentication, a confirmation flow to preview actions before finalizing them, daily email updates, and a robust action-integrity mechanism for auditing. However, it is important to note that while the application is fully operational, it has not been released to real end-users yet.
During its development, several engineering challenges were encountered, including silent integration issues and the unpredictability of third-party API responses. The lessons learned from these experiences highlight the importance of empirical validation over relying solely on metadata integrity in applied reinforcement learning systems that depend heavily on external live data sources.